Abstract

AUTOMATED 3D SEGMENTATION AND DETECTION OF PULMONARY NODULES IN CT SCANS USING DEEP LEARNING CONVOLUTIONAL NEURAL NETWORKS

Chaeyoung Lee (Tenafly High School)

Advisor: Blaine Rister (Stanford University)

Lung cancer is the leading cause of cancer mortality for both men and women worldwide. It is reported that early detection, by computed tomography (CT) screening, can decrease lung cancer mortality by 14-20 percent among high-risk populations. However, manual detection of nodules in CT images by radiologists is a very time-consuming and error prone process. Many diagnosis errors may be reduced with a second reader, but employing a second reader will be time- and cost- intensive. To address these problems, we developed a computer-aided detection (CAD) system that can automate nodule segmentation and detection by using convolutional neural networks (CNN). Our model uses a U-Net architecture to segment pulmonary nodules and a 3D CNN classification model to reduce false positives. The segmentation model alone achieved a sensitivity score of 0.71, but produced 7.76 false positives per scan (FPPS) with a detection accuracy of 0.18. After applying the false positive reduction model, the sensitivity score slightly dropped to 0.61, but the number of false positives significantly dropped to 1.54 FPPS from 7.76 FPPS and the accuracy increased to 0.75. Despite the slight drop in sensitivity, our CAD model still performs better or as good as a human radiologist who has an average sensitivity score of 0.49 and produces 0.6-2.1 FPPS. Based on the calculated sensitivity score and detection accuracy, the model shows potential to effectively aid radiologists in detection and diagnosis by reducing reading times or acting as a second reader.